Kling takes the lead, Flux 2 and Seedream 4.5 dazzle, and Runway Gen-4.5 released

All the news from the AI world for creators

Erik Knöbl

Dec 07, 2025

Good day, Directors

In today’s Newsletter:

1️⃣ Kling joins the leading pack (and brings Audio)

2️⃣ Black Forest Labs releases FLUX 2

3️⃣ Seedream 4.5: The new standard for consistency?

4️⃣ Runway Gen-4.5: The Director’s Tool

5️⃣ Quick news you may have missed (Deepseek, Mistral, OpenAI)

6️⃣ AI Videos of the Week

Let’s go.

We had two crazy weeks, and we are just starting December.

Kling joins the leading pack (and brings Audio)

In the current landscape of AI video, a sharp divide has emerged: tools that offer native dialogue and sound (like Veo3, Sora 2, and LTX-2), and those that are purely visual.

With its latest updates, Kling has firmly crossed that divide to join the leaders.

First, they have launched Kling O1, a true multimodal model that changes the workflow from “generating” to “directing.” It offers serious post-production power:

Context-Aware Editing: It can generate new shots with different angles and compositions while perfectly retaining the data from a reference video.
Element Control: You can remove or add specific elements within a video clip effortlessly. It brings the intuitive ease of image editing (think Nano Banana) into the complex world of video.

Why this matters: Simultaneously, Kling launched its 2.6 video model which finally supports native audio. This is a massive workflow unlock, allowing creators to generate usable, sound-synced clips in a single pass rather than juggling external sound tools.

Black Forest Labs releases FLUX 2

The open-source king is back. Black Forest Labs has released FLUX 2, and it is directly targeting professional workflows.

The new model features:

Extreme Realism: Pushing the boundary of “AI look” vs. photography. We’ll have to test that.
Massive Multi-Reference: You can now use up to 10 reference images to guide the generation. This is crucial for maintaining style or character identity, and puts Flux in direct competition with Nano Banana and Seedream 4.0.

Why this matters: They are releasing an open-source version. While proprietary models (like Midjourney) are powerful, FLUX 2 allows developers and studios to build their own custom pipelines without paying per-generation fees or risking data privacy.

Seedream 4.5: The new standard for consistency?

Bytedance has released Seedream 4.5, and unlike other updates that chase “more pixels,” this one chases refinement.

The update of version 4.0 focuses heavily on the biggest pain point in AI art: Subject Consistency.

Detail Fidelity: It understands the nuances of a style and applies them cohesively to similar images.
Spatial Reasoning: It actually understands the “location” and 3D structure of a scene, reducing weird geometry errors.
Fusion Power: Like Flux, it can generate images by fusing up to 10 different references.
Text & UI: It finally delivers clear, readable small text and precise facial rendering.

Why this matters: For brands and storytellers, “pretty” images aren’t enough; they need images that look like they belong to the same world. Seedream 4.5 appears to be the tool designed specifically for that cohesion.
Your move, Nano Banana.

Runway Gen-4.5: The Director’s Tool

Runway has introduced its new frontier model, Gen-4.5.

While other models are fighting over audio, Runway is doubling down on control. The model excels at understanding complex, sequenced instructions. You can now specify detailed camera choreography, intricate scene compositions, and precise timing of events, all within a single prompt.

The Catch: There is still no native audio. Buh.

Why this matters: Runway is positioning itself as the tool for filmmakers who want granular control over the visual physics and camera movement, even if it means handling the sound design elsewhere.

How do you feel about this bet? Are you using Runway lately?

Quick news you may have missed

The Chinese AI startup DeepSeek released its version 3.2. This open-source models rival OpenAI’s GPT-5 and Google’s Gemini 3 Pro in reasoning and agent tasks, using innovative sparse attention for faster processing of long texts at half the compute cost.
Mistral AI, a French startup, unveiled the Mistral 3 family: compact Ministral 3 models and the powerful Mistral Large 3. All are open-source, multilingual, multimodal, and optimized for NVIDIA hardware. They outperform larger rivals in coding and efficiency.
OpenAI CEO Sam Altman declared a “code red” in an internal memo, urging staff to prioritize ChatGPT improvements amid fierce competition from Google’s Gemini 3 (praised for top benchmarks) and rivals like Anthropic. OpenAI plans a new reasoning model soon to regain edge.

That’s all for now. Enjoy coffee, touch grass, create awesome videos.

Discussion about this post

Ready for more?